AI-Measured Tumor Volume on PSMA PET-CT Linked With Survival in Prostate Cancer

Tumor volume quantified by a fully automated artificial intelligence (AI) model on [¹⁸F]PSMA-1007 PET-CT, without any manual correction, was associated with overall survival (OS) among patients with prostate cancer undergoing imaging for primary staging or biochemical recurrence (BCR), according to results from a retrospective, single-center study published in Clinical Physiology and Functional Imaging.

Among the 267 patients included in the final analysis, 48 (18%) died during 5 years of follow-up. Whole-body total lesion volume (TLV) was significantly associated with OS in the combined cohort (P = .001). Compared with patients in the lowest tertile of whole-body TLV, those in the highest tertile had an age-adjusted hazard ratio (HR) for death of 3.20 (95% CI, 1.50-6.80); the HR for the intermediate tertile was 1.19 (95% CI, 0.49-2.86).

Among site-specific measurements, bone TLV was also significantly associated with OS (P = .016), with an HR of 2.20 (95% CI, 1.07-4.54) for high vs low bone TLV. The overall associations for prostate TLV (P = .063) and lymph node TLV (P = .071) did not reach statistical significance, although patients with high TLV at those sites had higher mortality than the low-TLV reference groups, with HRs of 2.23 (95% CI, 1.10-4.54) and 2.27 (95% CI, 1.10-4.66), respectively.

In the primary staging cohort, 29 patients (18%) died. Whole-body TLV (P = .014) and bone TLV (P = .016) were significantly associated with OS, with HRs for high vs low TLV of 3.99 (95% CI, 0.93-17.20) and 3.94 (95% CI, 1.30-11.93), respectively. Associations for prostate TLV (HR, 2.42; 95% CI, 0.98-6.01; P = .091) and lymph node TLV (HR, 2.34; 95% CI, 0.92-5.96; P = .180) were not significant. P values reflect overall likelihood-ratio tests across TLV categories.

In the BCR cohort, 19 patients (18%) died. The overall association between whole-body TLV and OS did not reach statistical significance (P = .061), although the HR for high vs low whole-body TLV was 3.47 (95% CI, 1.26-9.61). No significant overall associations were observed for prostate (P = .117), lymph node (P = .401), or bone TLV (P = .249) in this group.

"Importantly, TLV was quantified entirely automatically without manual correction, demonstrating that prognostically relevant information can be extracted from PSMA PET-CT examinations using AI. The fully automated approach makes large-scale volumetric assessment technically feasible without additional manual segmentation and may ultimately support applications such as risk stratification and treatment monitoring, although further validation is required before clinical implementation," lead author Johan Wittgren, MD, a physician in the Department of Clinical Physiology and Nuclear Medicine at Skåne University Hospital and a doctoral student at Lund University in Sweden, wrote in the publication with study coinvestigators.

Investigators retrospectively evaluated patients referred for [¹⁸F]PSMA-1007 PET-CT at Skåne University Hospital in Malmö and Lund between September 2019 and July 2021, who were consecutively invited to participate. Of the 267 examinations, 159 (60%) were performed for primary staging of newly diagnosed intermediate- to high-risk prostate cancer and 108 (40%) for BCR. The median patient age was 71 years (range, 42-87), and the median prostate-specific antigen (PSA) level was 13 μg/L (IQR, 7.6-22) in the primary staging group and 1.2 μg/L (IQR, 0.3-4.9) in the BCR group. In the primary staging cohort, 74% of patients had high-risk disease per D'Amico criteria; in the BCR cohort, 49% had intermediate-risk disease at diagnosis.

Scans were analyzed with a previously validated AI model developed by the research group, which has shown sensitivity similar to that of nuclear medicine physicians. None of the scans in this study were part of the model's training dataset. The model automatically segmented suspected lesions in the prostate or prostate bed, lymph nodes, and bone, and lesion volumes were summed into site-specific and whole-body TLV.

The primary outcome was OS. Site-specific TLV was categorized as no detectable tumor, low (at or below the median), or high (above the median), with the low group as reference; whole-body TLV was divided into tertiles. Mortality was compared using Cox proportional hazards models adjusted for age, and subanalyses were performed separately for the primary staging and BCR cohorts.

The investigators noted several limitations, including the retrospective, single-center design and a limited number of events, which may explain why the strength of the statistical evidence varied across anatomical sites and subgroups. Because the complete AI output was used without physician review, false-positive or nonspecific findings contributed to TLV. The model assigned a bone TLV above zero to approximately 60% of patients, whereas a prior study using manual delineation in a largely overlapping cohort identified bone lesions in approximately 20%. For that reason, the investigators cautioned that AI-derived bone TLV should not be interpreted as an exact measure of bone metastasis volume.

The findings were consistent with the group's earlier study using manual tumor delineation in a partially overlapping cohort, as well as with prior manually or semiautomatically derived tumor volume analyses. The authors also noted that a previous fully automated tumor volume study found higher mortality with increasing tumor volume among patients with metastatic castration-resistant prostate cancer receiving lutetium Lu 177–based therapy and said their results in earlier disease stages contribute to the generalizability of the approach. Future research should validate the results in larger cohorts with longer follow-up, explore optimized volume thresholds, and assess how AI-derived metrics can be integrated into clinical decision-making, they wrote.

Reference

Wittgren J, Ingvar J, Enqvist O, et al. Fully automated AI-based quantification of tumour volume on PSMA PET-CT images is associated with overall survival in prostate cancer. Clin Physiol Funct Imaging. 2026;46(6):e70104. doi:10.1111/cpf.70104